Physics-informed neural networks with residual/gradient-based adaptive sampling methods for solving PDEs with sharp solutions